Abstract
Accelerated MRI reconstruction must recover fine anatomical structure from undersampled Fourier measurements while maintaining reference fidelity. We study Posterior-Mean Rectified Flow (PMRF), which separates a distortion-oriented posterior-mean estimate from deterministic rectified-flow refinement. We first instantiate this decomposition for complex-valued accelerated MRI using a hard data-consistency projection on the posterior-mean anchor, yielding MRI-PMRF. We then introduce Endpoint-Neutral Guidance Rectified Flow (EN-GRF), which uses the acquisition as privileged training information to deform intermediate conditional paths while retaining their source and target endpoints. The guidance network is discarded at test time, so inference remains an eight-step ODE integration from the data-consistent anchor rather than posterior sampling. On 199 fastMRI single-coil knee volumes at R ∈ {4, 8, 12}, ENGRF reduces DISTS by 39–46% relative to the strongest evaluated deterministic baseline while remaining within 0.2 dB of the best PSNR. MRI-PMRF accounts for most of this gain, and ENGRF adds a further, statistically significant improvement in all feature distances at every acceleration without any additional inference-time conditioning. Our code is available at https://github.com/OmerTaub/ ENGRF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_055.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=QwrwgWdCYN
BibTex
@InProceedings{TauOme_Towards_MICCAISAT2026,
author = { Taub, Omer AND Ayzenberg, Lev AND Greenspan, Hayit},
title = { { Towards Distortion–Perception-Balanced MRI Reconstruction using Posterior-Mean Rectified Flow } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}